I'm testing the AMDGPU-PRO Beta driver (16.20.3) on Ubuntu 16.04 by running some OpenCL tests. The most significant issue I've noticed is that only ~680MB are available out of the 4GB of a Radeon R9-Nano. Given that only 1/4 of that can be allocated using a buffer object this proves quite restrictive.
In addition, I've observed some performance issues comparing with the fglrx 15.302.
Additionally, clinfo reports false info for compute unit count of R9 Nano (Fiji):
Max compute units: 14
That should be 64.
On a FX-7500 mobile APU the reported number of CUs is 7 where the correct one is 6 (Spectre device).
Using AMDGPU-PRO v.16.30.3-315407 on Ubuntu 16.04 64bit
With the latest amdgpu-pro driver (16.30.3-306809) the memory issue was resolved. However, the reported number of compute units is still 14! Weird.
I am seeing the same issue with the R9 380s...memory bug was fixed, but only showing 14 CUs, and clock is capped at 555mhz regardless of what is set in the card BIOS.
Clock is also reported to be 555MHz with Fiji. Power-play issue?